Cross-reference row cells are never parsed: parse_object receives the Submatch, not the cell #3542

Problem

parse_xref_row in polyfile/pdf.py recurses into the Position and Generation cells of a
cross-reference row by passing the Submatch it just created instead of the cell itself:

ret = Submatch("Position", b"", relative_offset=pos.pdf_offset - row_start, length=pos.pdf_bytes,
               parent=parent)
yield ret
yield from parse_object(ret, matcher=matcher, parent=ret, pdf_header_offset=pdf_header_offset)

The ObjectID cell two lines above passes obj_id, which is what the other three call sites of
parse_object do as well. A Submatch carries no pdf_offset, so parse_object falls through
every branch and yields nothing: the two calls are no-ops.

This predates #3464. PR #3545 moved the code into parse_xref_row verbatim rather than
changing it, because fixing it adds a PSInt submatch under every Position and Generation and
that is a change to PolyFile's output, not a guard.

Reproducer

Any PDF with a classic cross-reference table. For the two-object PDF in
tests/test_pdf.py:WELL_FORMED_PDF, the tree holds two Position and two Generation
submatches, and no PSInt beneath any of them.

Fix

Pass pos and gen_no. Confirm the four extra PSInt submatches per two-row table are wanted
before doing it — polyfile/templates/ renders the tree, and the JSON output format in
docs/json_format.md describes it.